key: cord-0307287-9yrhy69x
authors: Pupic, N.; Gabison, S.; Evans, G.; Fernie, G.; Dolatabadi, E.; Dutta, T.
title: Detecting Patient Position Using Bed-Reaction Forces and Monitoring Skin-Bed Interface Forces for Pressure Injury Prevention and Management
date: 2022-03-17
journal: nan
DOI: 10.1101/2022.03.15.22272323
sha: a3582d2d4cc1c0b19c11733554d5965bf6f8901d
doc_id: 307287
cord_uid: 9yrhy69x

Pressure injuries are largely preventable, yet they affect one in four Canadians across all healthcare settings. A key best practice to prevent and treat pressure injuries is to minimize prolonged tissue deformation by ensuring at-risk individuals are repositioned regularly (typically every 2 hours). However, adherence to repositioning is poor in clinical settings and expected to be even worse in homecare settings. Our team has designed a position detection system for home use that uses machine learning approaches to predict a patient's position in bed using data from load cells under the bed legs. The system predicts the patient's position as one of three position categories: left-side lying, right-side lying, or supine. The objectives of this project were to: i) determine if measuring ground truth patient position with an inertial measurement unit can improve our system accuracy (predicting left-side lying, right-side lying, or supine) ii) to determine the range of transverse pelvis angles (TPA) that fully offloaded each of the great trochanters and sacrum and iii) evaluate the potential benefit of being able to predict the individual's position with higher precision (classifying position into more than three categories) by taking into account a potential drop in prediction accuracy as well as the range of TPA for which the greater trochanters and sacrum were fully offloaded. Data from 18 participants was combined with previous data sets to train and evaluate classifiers to predict the participants' TPA using four different position bin sizes (~70{degrees}), 45{degrees}, ~30{degrees}, and 15{degrees}) and the effects of increasing precision on performance, where patients are left side-lying at -90{degrees}, right side-lying at 90{degrees} and supine at 0{degrees}). A leave-one-participant-out cross validation approach was used to select the best performing classifier, which was found to have an accuracy of 84.03% with an F1 score of 0.8399. Skin-bed interface forces were measured using force sensitive resistors placed on the greater trochanters and sacrum. Complete offloading for the sacrum was only achieved for the positions with TPA angles <-90{degrees} or >90{degrees}, indicating there was no benefit to predicting with greater precision than with three categories: left, right, and supine.

Our team has designed a position detection system for home use that uses machine 17 learning approaches to predict a patient's position in bed using data from load cells 18

under the bed legs. The system predicts the patient's position as one of three position 19

categories: left-side lying, right-side lying, or supine. The objectives of this project were 20

to: i) determine if measuring ground truth patient position with an inertial measurement 21 unit can improve our system accuracy (predicting left-side lying, right-side lying, or 22 supine) ii) to determine the range of transverse pelvis angles (TPA) that fully offloaded 23 each of the great trochanters and sacrum and iii) evaluate the potential benefit of being 24 able to predict the individual's position with higher precision (classifying position into 25 more than three categories) by taking into account a potential drop in prediction 26 accuracy as well as the range of TPA for which the greater trochanters and sacrum 27

were fully offloaded. 28

Data from 18 participants was combined with previous data sets to train and evaluate 29 classifiers to predict the participants' TPA using four different position bin sizes (~70°, 30 45°, ~30°, and 15°) and the effects of increasing precision on performance, where 31

patients are left side-lying at -90°, right side-lying at 90° and supine at 0°). A leave-one-32 participant-out cross validation approach was used to select the best performing 33 classifier, which was found to have an accuracy of 84.03% with an F1 score of 0.8399. 34

Skin-bed interface forces were measured using force sensitive resistors placed on the 35 greater trochanters and sacrum. Complete offloading for the sacrum was only achieved 36

for the positions with TPA angles <-90° or >90°, indicating there was no benefit to 37

predicting with greater precision than with three categories: left, right, and supine. 38 39

Keywords 40

Pressure injuries, bed sores, machine learning, neural networks, prevention, healing, 41

repositioning, skin-bed interface 42 43 1 Introduction supine, and right-side lying were represented by -90°, 0°, and 90°, respectively. Figure 1  89 shows a visual representation of the TPA. adjusted to avoid the bin boundaries coinciding with the positions that participants were 103 asked to adopt. 104 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. A convenience sample of 20 healthy participants (10 males, 10 females) was recruited 110

for this study. Able-bodied participants were included with no existing pressure injuries. 111

All participants provided their informed consent, and the study protocol was reviewed by 112

the one around the chest, one around the pelvis, and one around the arm in order to 131 collect ground truth data for the sternal angle, pelvic angle, and heart rate, respectively. 132 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The data was collected in two phases: a) the primary phase, where participants were 146

instructed to cycle through a series of 11 unique positions at 0°, ±15°, ±30°, ±45°, ±60°, 147 and ± 90°; and b) the random phase, where participants could assume any position they 148

wanted to from -180° to +180° to account for the wide range of positions that can be 149 adopted in bed. In addition, participants were asked to assume one prone position.

All primary and random phase positions were held for three minutes. 

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) Load Cell Data -Load cell signals were exported from Netforce and processed offline 172

using MATLAB 2020a. The data was manually segmented into trials by removing 173

sections where the participants were changing positions. Next, the center of mass of the 174 bed-patient system was calculated using equations 1 and 2 below where CoM_x and 175

CoM_y refer to the center of mass in the x (parallel to the short axis or width of the bed) 176

and y (parallel to the long axis or length of the bed) directions, respectively. The data 177

processing will be performed in the same manner as the study by Wong et al (12) .

Below is an explanation of how the data processing was executed, where LH 

198 199

Finally, components of the signal that captured changes resulting from the cardiac cycle 200

(rmsPulse) were isolated. MATLAB's filtfilt function was used to bandpass filter the sum 201 of the LH and RH signals using a personalized equiripple finite impulse response filter. 202 203 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) home environment and classifications performed by 3 independent and blinded raters). 215

The IMU data was further classified four more times using the generated Euler angles, 216 once for each of the different bin sizes, to allow for more precise TPA detection. 217 218

FSR Data -Only FSR data from the primary phase was used for this analysis as the 219

positions were consistent between all participants. The FSR data was manually 220 annotated by the author to assign position codes using the video data as ground truth 221 guidance. their in-lab participant study were used. These features were extracted from the load 245 cell data (Table 1) . 246 247 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) the new data to predict either supine, left, or right. After this classification, the new data 253 was divided into three separate smaller data sets based on its IMU ground truth label 254

(either supine, left, or right). The data from each of the three positions was fed into its 255 respective model for Level 2 classification, where the model specified a more precise 256 bin of angles that each position belonged to. Level 2 classification was repeated four 257 times, once for each bin size (15°, ~30°, 45°, ~70°). 258 259 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. 

Leave-One-Participant-Out -A leave-one-participant-out cross validation approach was 267 used to evaluate the accuracy of the classifier, while maximizing the number of training 268

observations. Using this method, a classifier was trained on a data set that incorporated 269

17 participants and tested on the one excluded participant. This procedure was 270 repeated 18 times, once for each participant. The overall performance measures were 271 estimated from the averaged errors for each individual test sample.

Incremental Learning -Incremental learning was used to evaluate the potential of the 274 classifier to adapt to the left-out participant. The classifier was trained using different 275 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) Machine Learning Classifiers - Table 3 shows a list of models used in both Level 1 and 280 2 classifications. For Level 1, both machine and deep learning models were used. For 281

Level 2, only machine learning models were used as there was not enough data to 282

warrant the use of deep learning. 283 284 Incremental learning levels were also compared for each of the top models to determine 321 its impact on performance. Each incremental learning level was compared to its 322 adjacent value(s) (i.e., 0% to 10%, 10% to 20%, and 20% to 30%).

Offloading Data -The percentage of maximum load recorded in each position was 325 compared for each of the three FSRs. The statistical analyses were performed twice for 326 each sensor, for a total of six analyses, to compare all adjacent positions from 90° to 0° 327 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) Descriptive statistics of the 20 participants recruited for this study are provided in Table  339 9. 340 341 and right for each incremental learning level. Since the ILL of 30% performed best, we 348 conducted our analyses on these models. 349 350 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted March 17, 2022. ; https://doi.org/10.1101/2022.03.15.22272323 doi: medRxiv preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. For the left and right trochanters, only the positions from -90° to 0° and 0° to 90°, 394 respectively, were assessed. This decision was made because a trochanter is 395 completely offloaded when a participant is on the opposite side, meaning the force 396

reading would be 0. The above does not hold true for the sacrum, so it was assessed 397

for the entire range of positions from -90° to 90°.

Left Trochanter -The left trochanter loading data for positions -90° to 0° was analyzed 400

and confirmed to be non-parametric. A Friedman's ANOVA was significant, χ2 (5) 7b . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Figure 9 shows the percentage of max load felt at the right trochanter as participants . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. 

Sacrum -The sacral loading data for positions 90° to 0° was analyzed and confirmed to 436 be non-parametric. A Friedman's ANOVA was significant, χ2(5) = 75.218, p < 1x10 -14 .

Multiple post-hoc Wilcoxon Rank Sum tests with Bonferroni corrections were used to 439 compare adjacent primary positions to determine if there was a difference in percentage 440 of maximum load. In total, five comparisons were made, changing the p-value needed 441

to reach significance to p < 0.01. Four of the comparisons were statistically significant, 442 90° to 60°: V = 7, p < 0.001; 60° to 45°: V = 7, p < 0.001; 45° to 30°: V = 55; p = 0.33; 443 30° to 15°: V = 4, p < 0.001; 15° to 0°: V = 9, p < 0.001.

The sacral loading data for positions -90° to 0° was analyzed and confirmed to be non-446 parametric. A Friedman's ANOVA was significant, χ2(5) = 67.679, p < 1x10 -12 . . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. 

Tables 12 and 13 show the overall mean accuracy and F1 scores with their respective 464

standard deviation values across all 18 participants from Level 2 left and right 465 classification using an ILL of 30%. Only the top two models are shown. 466 467 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The best performing model was the XGBoost model with a mean accuracy of 84.03% ± 503

12.17%, and mean F1 score of 0.8399 ± 0.1226. This accuracy is an improvement over 504 the ~68% that Wong et al.'s previous model was able to achieve on this data. It is likely 505 that this accuracy represents an underestimate of the actual accuracy of our new 506 system as the training set includes data without the IMU ground truth data. 507 508

The deep learning methods generally performed worse than the machine learning 509 models, except for the MLP 2 model which performed at a comparable level to the LGB. 510 However, this model had an overly complicated architecture for the amount of data it 511 was processing, so it may have overfit to the data. In general, the most likely reason for 512 the poor performance of the CNN and RNN was the lack of data as deep learning 513

traditionally relies on very large data sets.

Comparing Machine Learning Models -The omnibus test found an overall significant 516 difference between the performance of all models and post-hoc comparisons of the top 517 three models found statistically significant differences. In terms of mean accuracy and 518 F1 score, the XGB model performed significantly better than both the LGB and MLP 2 519 models. For mean accuracy, the LGB and MLP 2 were found to have performed 520

comparably, whereas for mean F1 score, the LGB model outperformed the MLP 2 521 model. Therefore, the XGB model was statistically significantly better than the other 522 models and should thus be included in future work when testing position prediction.

Comparing . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) to -30° and -15° to 0° almost met significance. This finding indicates that rotating a 548 participant from a more extreme position to a less extreme position between the 549 positions -60° to -45° and -30° to -15° will result in significant offloading of the left 550 trochanter compared to the previous position.

Right Trochanter -The right trochanter was loaded for most angles between 0° to 90° 553 (as shown in Figure 9 ). A Friedman's ANOVA indicated that there was a significant 554 difference between the mean percentage of maximum force experienced on the right 555 trochanter between different right-side lying positions. Post-hoc comparisons compared 556 adjacent right positions and identified that the percentage of maximum force on the right 557 trochanter was significantly different for all adjacent positions except for 90° to 60°. This 558 finding indicates that rotating a participant from a more extreme position to a less 559 extreme position between the positions 60° to 45°, 45° to 30°, 30° to 15°, and 15° to 0° 560 will result in significant offloading of the right trochanter compared to the previous 561 position.

Sacrum -The sacrum was only completely offloaded at -90° and 90° (as shown in 564 Figure 10 ). A Friedman's ANOVA reported a significant difference between the mean 565 percentage of maximum force experienced on the sacrum between different right-side 566 lying positions. Post-hoc comparisons compared adjacent right positions and identified 567 that the percentage of maximum force on the sacrum was significantly different for all 568 adjacent positions except for 45° to 30°. This finding indicated that rotating a participant 569 from a less extreme position to a more extreme position between the positions 0° to 15°, 570

15° to 30°, 45° to 60°, and 60° to 90° resulted in significant offloading of the sacrum 571 compared to the previous position. 572 573 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted March 17, 2022. ; https://doi.org/10.1101/2022.03.15.22272323 doi: medRxiv preprint A Friedman's ANOVA reported a significant difference between the mean percentage of 574 maximum force experienced on the sacrum between different left-side lying positions. 575

Post-hoc comparisons were used to compare adjacent left positions and they identified 576 that the percentage of maximum force on the sacrum was significantly different for -90° 577

to -60°, -30° to -15°, and -15° to 0°. Changes from -45° to -30° almost met significance. 578

This finding indicates that rotating a participant from a less extreme position to a more 579 extreme position between the positions 0° to -15°, -15° to -30°, and -60° to -90° will 580 result in significant offloading of the sacrum compared to the previous position.

Optimal Offloading -The trochanter opposite to the side a patient was turned on will be 583 completely offloaded, making the trochanters easier to offload than the sacrum. It 584 appeared that the sacrum was not fully offloaded in any position that required the use of 585 a support pillow as the patient's sacrum was likely pressed up against it. If complete 586

offloading is required for adequate tissue healing, it may be necessary for patients to 587 assume a side-lying position that can be maintained without the use of assistive device 588

to maintain the position. If assistive devices are needed, it may be important to ensure 589

they have a cut out around the sacral area to ensure it is being properly offloaded. 590 591

Optimal Bin Size -The results indicated that the smallest bin size needed to detect 592 meaningful changes in offloading is 15°. However, if patients require complete 593 offloading to heal, then classifying positions as supine, left, or right will suffice. As such, 594

it may be more important to focus on accurately detecting large positional changes like 595 those in Level 1 classification to ensure offloading is occurring on a scheduled basis. 596

More precise detection may be useful in recognizing smaller self-repositioning efforts 597

and determining their impact on high-risk areas. 598 599

Level 2 classification separates the right, left, and supine classifications from Level 1 601 classification into smaller bins. Tables 12 and 13 summarize the results from the top two 602 models for Level 2 classification of left and right bins based on bin size. The best 603

performing models vary depending on bin size, but the XGB model was best in six out 604 of eight cases for mean accuracy and seven out of eight for mean F1 score . The table  605 also shows a trend of left-side positions being predicted correctly more often than right-606 side positions. The reason for this finding is currently unclear. 607 608

Comparing Bin Sizes -Tables 12 and 13 and Figures 11 and 12 show the effect of bin 609 size on prediction accuracy. The results show that the accuracy of predictions 610 decreases as the precision, or number of bins, increases. 611 612

ANOVAs indicated that the only significant difference in bin sizes was in the Level 2 left 613 mean accuracy comparison. When further analyzed, post-hoc t-tests indicated that the 614 45° vs 30° and 30° vs 15° bin comparisons were significant.

These results are important because they indicate that there is likely a trade-off 617 between accuracy and precision when making predictions. It will be important to 618 optimize the bin size for this system to ensure it is recording and classifying movements 619 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted March 17, 2022. ; https://doi.org/10.1101/2022.03.15.22272323 doi: medRxiv preprint

bin size should be optimized based on information gathered 620 from offloading data and clinical expertise to decide what is the smallest positional 621 change that needs to be captured

Study Limitations 624 This study included a number of technical and clinical limitations as described below: 625 626 The technical limitations of this study included: 627 1. Load cell data was filtered with customized parameters for each participant, 628 which may have led to an increase in classification accuracy compared to using 629 generic parameters

Most of the training data did not contain IM ground truth, so it could have 631 negatively impacted the accuracy

The data set was too small to run one-shot learning to compare its accuracy with 633 the current hierarchical approach

The neural network architectures selected by hyperparameter tuning were very 635 complicated and may have overfit the data

Level 2 classification had an imbalance of positions that were >90° and <-90°, 637 which could have negatively impacted the prediction accuracy for more extreme 638 positions. 639 640 The clinical limitations of this study included: 641 1. The patient population was predominantly young, healthy individuals, which likely 642 did not reflect the performance with a population of older adults with/at-risk of PI. 643 2. Certain primary positions participants were asked to adopt were unnatural, which 644 may have impacted the participants' abilities to relax/breathe normally. 645 3. Individuals were supported using pillows that were occasionally placed against 646 the sacrum in side-lying positions, which could have resulted in overestimates in 647 sacral load

The way in which the IMUs were attached could have been more reliable to 650 ensure they did not move while patients were changing positions

The FSRs were placed by participants under guidance of the author, which 652 means they may not have been placed in the correct anatomical position every 653 time

The FSRs occasionally fell off the participants during the study and needed to be 655 re-attached mid study. 656 657 4.5 Future Work 658 Future work will include: 659 1. Collecting overnight data from at-risk patients in their own homes. 660 2. Revising pillow placement during offloading and trying out different repositioning 661 aids to see the impacts on sacral loading

Validate a safe way to use IMUs for overnight data collection of at-risk patients. 663 4. Treating the detection of patient position as a regression task instead of a 664 classification task to evaluate performance

An IMU mounted to the pelvis improved position detection accuracy for supine, 669 left, or right from ~70% in our previous work to 84.2% ± 11.8% for the best 670 performing model

The right and left trochanters were completely offloaded for TPAs of 0° to 90° 672 and 0° to -90°, respectively. The sacrum was only completely offloaded for TPAs 673 of >=90° and <=-90°, highlighting a potential limitation of the existing clinical 674 guidelines suggesting individuals be rotated between TPAs of -40° and 40°

Prediction accuracy decreased as the precision increased

Pressure ulcers: Updated guidelines for treatment and prevention

Best practice 681 recommendations for the prevention and management of pressure injuries. 682 Wounds Canada

What is new in our understanding of pressure injuries : the 684 inextricable association between sustained tissue deformations and pain and the 685 role of the support surface

How patient migration in bed affects the sacral 687 soft tissue loading and thereby the risk for a hospital-acquired pressure injury

The future of pressure ulcer prevention is here: Detecting and 690 targeting inflammation early

Revised 695 National Pressure Ulcer Advisory Panel Pressure Injury Staging System

Quality of 698 Care for Hospitalized Medicare Patients at Risk for Pressure Ulcers

Body positioning of 701 intensive care patients: Clinical practice versus standards

A novel 708 system to tackle hospital acquired pressure ulcers

Identifying gaps, barriers, and solutions in 713 implementing pressure ulcer prevention programs

The national cost of hospital-acquired pressure 717 injuries in the United States

Toward 719 mitigating pressure injuries: Detecting patient orientation from vertical bed 720 reaction forces

Measuring 724 Repositioning in Home Care for Pressure Injury Prevention and Management

Classification of lying position using load cells 727 under the bed

Evaluation and 729 accurate diagnoses of pediatric diseases using artificial intelligence